Severe class imbalance: failures (pos) are ~1.7% of the data. Heavy missing values (the raw data uses na markers). Train / validation / test split — the training data is split 80/20 (stratified) ...